An agent could be viewed as a function that takes the perceived universe P and the universe of actions A.
With function agent, P partially determines A.
Partially because the agent can potentially ignore part of P.
P should be a sequence of perceptions. We map it to one action at a time.
TODO: Rewrite with formula.
We should design the performance based on how well the model performs, not on how we think hit should achieve the result.
For each possible percept sequence, a rational agent should select an action that is expected to maximise its performance measure.
Ingredients:
Rationality is not omniscience tho, we are learning to extend prior knowledge, or else the agent would lack autonomy???
if status == Dorty:
return Suck
else if location == A
return Right
else if status == B
return Left
It requires a transition model and a sensor model. Idk man...
Keeps track of the world state and a set of goals that it's trying to achieve. Then it chooses an action that will eventually lead to the achievement of its goals.
This is done by basically understanding what its action do, predicting the outcome of the action and then actually acting on it to achieve the goal.
To avoid conflicting goals, we have a performance measure based on utility. It will choose the action that is expected to have the best utility.
The system adapts to the environment. There is a performance assessment to know if the learning is working or not, the learning agent actually changes the performance agent to try and improve its performance.